AssemblyAI and Cartesia split voice agents by layer

AssemblyAI and Cartesia are competing on more than voice quality, with each company now offering a broader stack for voice agents rather than a single text-to-speech model. The comparison turns on three separate layers: speech synthesis, speech recognition and agent orchestration, along with pricing, language coverage and published benchmark results.

Evidence and context

According to the companies' pricing pages and docs as of October 28, 2026, AssemblyAI lists AssemblyAI-TTS-1 for text-to-speech, Universal-3.6 Pro Realtime for streaming speech-to-text and LLM Gateway for model access, while Cartesia lists Sonic 3.6 for text-to-speech, Ink 2 for speech-to-text and Managed Agents as a hosted agent product billed per minute. The source comparison says the products are intended to be used in different ways: AssemblyAI's components are designed to run inside LiveKit, Pipecat or a user's own orchestrator, while Cartesia includes hosted agent tooling.

On text-to-speech, AssemblyAI says AssemblyAI-TTS-1 was trained on millions of hours of audio, including real conversational data, and offers thousands of voices across six languages: English, Spanish, French, German, Italian and Portuguese. The announcement also says the product supports instant voice cloning in the dashboard, three interfaces and unlimited concurrency. Cartesia's Sonic 3.6 is described in its documentation as generally available with 44 languages, and the comparison notes that Cartesia publishes dated snapshots such as sonic-3.6-2026-08-27 alongside a moving sonic-3.6 alias.

The source material also draws a line between voice tuning controls. Cartesia's docs and tooling include pronunciation dictionaries, while AssemblyAI's TTS connection parameters do not include speed, pitch, emotion, style or stability controls and do not support SSML. Instead, the AssemblyAI guide says users shape delivery through punctuation, capitalization and written-in filler words.

For telephony, the two services appear aligned. Both output pcm_mulaw and pcm_alaw at 8 kHz for carrier networks, and both documents reference Twilio's 8 kHz μ-law format as a target.

Published evaluation figures in the source material favor different models depending on the metric. In AssemblyAI's evals from the launch post, entity accuracy was 93.7% for AssemblyAI-TTS-1 versus 82.5% for Sonic 3.6, while median time to first byte was 47 ms versus 143 ms. The launch post's production time-to-first-audio figure for AssemblyAI-TTS-1 was 90 ms. Predicted listener preference, measured as an automated MOS estimate on a 1-to-5 scale, was close at 3.85 for AssemblyAI-TTS-1 and 3.82 for Sonic 3.6.

The comparison also cites Coval's TTS leaderboard, where AssemblyAI-TTS-1 recorded 1.7% word error rate and Sonic 3.6 recorded 1.8% as of October 28, 2026. The source notes that Coval re-measures daily over a rolling 30 days, so the ranking can change. It also points out that latency figures are not always directly comparable because they can measure different milestones, such as time to first byte versus time to first audible audio.

On speech recognition, AssemblyAI's Universal-3.6 Pro Realtime is described as a streaming flagship with 32 languages, built-in turn detection and pricing of $0.45 per hour. Cartesia's current streaming model is Ink 2, released on September 17, 2026, with support for English, French, Hindi, Japanese and Spanish. The source material says Cartesia's earlier Ink-Whisper model appears under "Older Models."

AssemblyAI's benchmarks page, as quoted in the material, shows a clear split between its own testing and third-party tables. In its own suites, the page says Universal-3.6 Pro Realtime was tested against Cartesia Ink-Whisper, not Ink 2, and posted a mean WER of 5.19% versus 15.17% on the English voice-agent benchmark, along with entity error rates of 14.4% versus 40.1%. The same page says Ink 2 appears in the systems table without a current run on those suites, so there is no like-for-like Ink 2 row from AssemblyAI's own harness yet.

The page also reproduces third-party tables that include Ink 2. In the Cekura Converse-STT clean-clips benchmark, Universal-3.6 Pro Realtime posted 1.77% WER and Ink 2 posted 2.39%. In the Cekura real-calls benchmark, the source lists 4.50% for Universal-3.6 Pro Realtime and 3.09% for Ink 2. It also lists median time to final text of 91 ms for Universal-3.6 Pro Realtime and 123 ms for Ink 2. The material does not say these tables were produced on the same date or under identical test conditions, so they should not be treated as a single unified ranking.

Pricing and concurrency limits also differ across the stacks. The source material lists AssemblyAI's TTS price at $15 per 1 million characters with no commitment, and its STT price at $0.45 per hour. Cartesia's TTS is described as costing about 1 credit per character under monthly credit plans, while Ink 2 is priced at 3 credits per second of audio. For concurrency, AssemblyAI's TTS is described as unlimited, with new sessions per minute rate-limited, while Cartesia publishes plan-based concurrent request caps from 2 on Free to 15 on Scale, with custom limits for Enterprise. For STT, AssemblyAI is described as offering unlimited concurrency with new-session rate limits, while Cartesia lists 8 concurrent requests on Free up to 60 on Scale, plus custom Enterprise limits.

The source material does not establish a single winner across the full voice-agent stack. It instead shows two products that overlap in some areas and diverge in others, with AssemblyAI emphasizing agent-oriented TTS and streaming STT components and Cartesia combining TTS, STT and hosted agents. The timing of the next verified product update or benchmark refresh is not disclosed in the material.